AI generated parody lyrics – BSc thesis
Title: AI generated parody lyrics – How are the lyrics that the software Weird A.I. Yankovic
produces regarded and how can they be improved?
Title in Swedish: AI-genererade parodiska låttexter – Hur upplevs parodilåttexterna som programmet Weird A.I. Yankovic maskingenererar och hur kan de förbättras?
Author: Daniel Andersson
Institution: Stocholm University (SU), Department of Computer and Systems Sciences (DSV), Sweden
Date and language: Autumn term 2021, Swedish
Type: Thesis at BSc level degree project 15 HP (sv: Självständigt arbete på grundnivå [kandidatexamen])
Key words: Language technology, Parody lyrics, Perceived funniness, AI, Neural networks
Advisor: Mirjam Palosaari Eladhari
Abstract
This essay concerns AI generation of parody lyrics, based on an existing program called Weird A.I. Yankovic (hereafter WAIY) that is developed by Mark O. Riedl at Georgia Institute of Technology. It is a Python program that loads the powerful language models GPT-2 and XLNet and takes a short parody context description as well as an existing song text as input. While the syllable and rhyme structure are retained, the words are replaced with newly generated text based on the entered context. The result is new lyrics that can be sung to the original melody but are intendent to gain comical qualities given the new content.
To get an overview of present and potential functionality, it has been investigated what approaches and functionality other AI programs that can generate lyrics use, and which of these could be implemented in WAIY. One of the identified potentials for improvement lies in additional training of the language models that generate the text content, as these are only generically trained in the original version. Accordingly, the main research question asks how the lyrics are affected by more specialized training of the language models.
In order to answer the research question, action research was performed where WAIY was further developed with new training of the GPT-2 language model on a few different specialized databases which are then evaluated. The evaluation concluded that a database of parody encyclopedia articles from Uncyclopedia gave the best results. The complete implementation of this version is called WAIY+.
To investigate whether WAIY+ is perceived as generating funnier lyrics, a survey was conducted in which participants were told to rate a number of lyrics’ funniness and realness, in addition to answering background questions. The results of the survey showed that a significant difference was measured where lyrics generated by WAIY+ were perceived as more realistic than those of generated by the original version. WAIY+’s lyrics were also perceived funnier even though this difference was not statistically significant.
The survey also showed that a piece of real lyrics written by a famous comedy singer artist was regarded as considerably more fun and more realistic than any of the AI-generated lyrics. This indicates a great potential for future progress in the field, and several proposals for further research have been presented in this study.
Through the survey, a few differences between WAIY and WAIY+ have been confirmed and these indicate that some improvement can be achieved through the proposed implementation. The added code of WAIY+ is presented in its entirety in the essay to enable further work in the research field.